A two-stage training method improves GUI grounding

When an instruction points to a small or unfamiliar control, how can training help a model learn from misses when every sampled click may be wrong? The paper asks whether richer guidance before reinforcement learning can address that gap.

D2RL first guides a model’s own click attempts with a teacher, then trains it on whether each click lands inside the requested element.

In an ablation, the authors compared post-training variants from the same supervised-trained checkpoint using the same post-training data.

Across five GUI tests that score whether a predicted click lands in the target area, D2RL averaged 69.9 points, versus 68.0 for supervised training alone, 69.1 with OPSD alone, and 68.9 with RLOO alone.

A control that used the post-training examples during supervised training before RLOO averaged 69.1 points. The authors interpret D2RL’s 0.8-point gap over that control as evidence that its result is not simply due to extra data exposure.

The results cover five GUI grounding tests and the reported training setups. The RLOO† comparison does not establish which procedural difference accounts for its gap with D2RL, because the exact starting checkpoint and other training conditions are unspecified.